Lend and borrow resources across teams with AWS Batch’s quota management for SageMaker Training jobs

This article introduces quota management, a new scheduling capability in AWS Batch that enables administrators to allocate and share GPU resources across teams running SageMaker Training jobs.

• Divide compute capacity into quota shares organized by team, project, or workload with configurable resource-sharing strategies
• Three sharing strategies: RESERVE (no sharing), LEND (share idle capacity), and LEND_AND_BORROW (flexible sharing)
• Cross-share preemption automatically reclaims lent capacity when the lending quota share receives new jobs
• In-share preemption allows higher-priority jobs to preempt lower-priority running jobs within the same quota share
• Integrate with SageMaker Python SDK v3 to submit training jobs to quota-management-enabled job queues
• UpdateServiceJob API enables dynamic priority adjustments for already-submitted jobs

Quota management helps organizations maximize GPU utilization, reduce infrastructure coordination overhead, and enable teams to focus on workloads rather than resource management.